Cartesia shipped Sonic-3.6 this week, and for once the flex is independently verifiable. The text-to-speech model now sits at #1 on both Artificial Analysis speech leaderboards — 1,283 Elo on the Provider Voice board and 1,123 on the Controlled Voice board. That second number is the one worth paying attention to. The Controlled board clones every model onto the same eight reference voices, stripping away any advantage from a flashy voice catalog and isolating the synthesis engine itself. Sonic-3.6 leads there too, with Sonic-3.5 in second and ElevenLabs Eleven v3 in third.
What changed? Cartesia says naturalness. The launch page frames the usual speed-versus-quality tradeoffs as architectural rather than inevitable — Sonic runs on state space models instead of transformers, and Cartesia claims sub-90ms time-to-first-audio. That’s vendor-stated model latency, not a measured end-to-end round trip, so anyone deploying this should run their own benchmarks before trusting the number.
The model is available now in beta through Cartesia’s hosted API. There are no open weights, no Hugging Face repo, no self-hosted option. You rent it. Pricing normalizes to $49 per 1M characters on Artificial Analysis — exactly half of ElevenLabs Eleven v3 at $100, though well above Speechify Simba 3.2 at $10 for a model that scores 1,240 Elo. Cartesia actually sells credits, not characters, with the $299 Scale tier covering roughly 10,667 TTS minutes and 15 concurrent requests. Voice agents bill separately at $0.06 per minute.
The production feature set is clearly aimed at agent transcripts rather than audiobook narration. Inline expression tags let you drop [laughter] directly into text. Voice cloning takes about 10 seconds of audio. There’s IPA support for pronunciation overrides — the kind of thing that matters when your IVR needs to say “subpoena” correctly. And alphanumerics read natively, so order numbers and confirmation codes don’t require preprocessing. The launch demos show English with natural pauses and filler words, plus Hinglish code-switching between Hindi and English.
Here’s the thing about TTS leaderboards: they’ve been volatile. Sonic-3.5 held the top spot three months ago, and now it’s been bumped by its own successor. That’s healthy competition, but it also means buyers should evaluate models against their specific use case rather than chasing Elo. Cartesia’s docs still list Sonic 3.5 as stable, and partners carry it — a reminder that beta means beta, even when the leaderboard says you’re #1.
